I am currently finishing up my writing my trainer class which can handle re-training by layer, prediction, some explainability and grid search incorporation with logging(all inspired by fastai). I am relatively new to this field(<8 months) and decided to do this to speed up my work on new problem statements. I am curious to know whether it is general practice or you would rather spend the time to write everything from scratch?
I would say it depends on your use case and the availability of the needed functionalities.
Often you’ll find the wanted features in higher level APIs like fastai, Ignite or Lightning.
These APIs provide you a different layer of abstraction so that you can chose the appropriate one.
However, if you’ve already a working code and just need to apply some form of logging, it’s sometimes more convenient (for me at least) to write this functionality myself.